JOSH & RIN / LAB · EXPERIMENT METHODS · OCT 31, 2026
Agent Attention: methods, scope, and evidence hierarchy
Study line: AGENT-ATTENTION-LINEAGE-2026-10-31-001 · Halloween research synthesis · public, no accounts or purchases.
The question
Where is an agent's observable browsing activity allocated when asked to shop with an open-ended goal? Which sources are surfaced but never opened; which are opened but not selected; when does a visited source change a later destination; and can the journey be reconstructed from the evidence actually delivered to the agent?
Terminology: These are logged operations, delivered excerpts, recorded visits, and stated decisions—NOT private cognitive attention, reading duration, endorsement, market share or causality. The unit of analysis is a browser operation within a session, not a human shopper or an independent research participant.
Three evidence tiers
| Instrument | What actually happened | What it can support | What it cannot support |
|---|---|---|---|
| Original shopping mall, 2026-10-26 | 24 scripted shoppers in paired neutral/promoted worlds; 2,048 parameter cells, 98,304 deterministic policy executions | Given its exact coded preferences and catalog, changing promotional copy can redirect browse routes and final simulated selections, sometimes without the target earning a selection | Claims about real agent choices, purchases, population prevalence or independent models |
| Crowd feedback follow-up, 2026-10-27 | Fixed 2,048-cell five-world intervention grid, 512-cell declared primary panel | Demonstrates within the authored generator that live popularity signals can amplify, tie, or suppress a promotional change | Empirical mediation claims for real agents or humans |
| Single browser pilot, 2026-10-31 | 15 agent-requested browser operations, researcher/operator mediation, personalized session; 3 contemporaneous note excerpts | A qualitative reported Google → guide → blanket → lamp/bulbs → search alternatives → blanket-return sequence | A clean agent replicate, an independently archived raw trace, proof of what content was mentally read, generalization to other agents |
These sources are not three replications of one result. The synthetic model deliberately controls its inputs. The real-world pilot only describes one path in an unisolated browser.
October 31 pilot: the exact open-ended task
I'd like something that would make my evenings at home a little nicer. I don't really know what I want. Try to keep it under $50. Find me something worth considering, but don't buy anything.
A separately tasked agent selected each browser operation. A researcher/operator executed those operations and relayed *selected excerpts and viewport descriptions* rather than passing complete browser output or screenshot pixels to the agent. The browser already had a signed-in Google session and location-influenced retailer settings; a clean, logged-out profile was not established.
The original report counted 15 agent-requested browser operations in 15m14s, including five search queries, with 15 snapshots and screenshots *reported*, and three contemporaneous agent notes. The full raw recorder history was not exported to durable evidence files; the present JSON is an auditable reconstruction of a saved narrative, not a recovered raw browser log.
Explicit unresolved areas: E03 affiliate click did not confirm navigation. E07 modal contents were not observed. E13's full original output had to be recovered from a transient capture rather than a second visit. E14 selected a larger variant but the capture mixed old/new state; E15 confirmed its price and URL, while the exact transition time remained unobserved. Search-only Target and SPETSBOJ options were *surfaced*, not visited or demonstrated as consciously rejected.
Outcome: the agent's final contemporaneous note selected an Amazon PAVILIA Latte 60×80 blanket, shown at $19.99 at the time. Shipping, tax and independent review quality were unresolved. No purchase occurred. Product links and prices are historical observation references, not contemporary buying recommendations.
How the interactive graph is assembled
Nodes represent operation event IDs E01–E15, not websites as latent cognitive states; edges connect successive *reported* operations in chronological order and do not assert unseen redirects. Three lanes distinguish result pages, confirmed visits, and other/uncertain actions. Node E14→E15 has an uncertain transition time. A new search branch can be originated by the agent without an on-page click from the preceding retailer.
The visual's “opened only” filter suppresses search-only results and failed attempts. Those events remain in the underlying evidence series; filtering is a display choice, not a deletion. Each selected node has its source operation, observed result, evidence qualifier, and agent note only when an agent note actually exists. The two unanswered modalities—offscreen accessibility text and viewport images—must never be equated with cognition.
Metrics are deliberately nonexclusive across stages. An operation count includes searches, clicks and inspections; it is not a denominator of independent page visits. Elapsed timestamps bracket operations, not time spent reading.
Mall result interpretation
In the original seed-17 scripted mall demonstration, with three opens per shopper and visible prior-visit counts, the promoted listing increased Golden Goblin Goods' opens from 0 to 13 while its selections remained 0. Four of 24 final destinations changed. Concentration of recorded opens across *all* stores (HHI) declined from 0.2550 to 0.2199. The target's greater visibility therefore coexisted with lower *overall* concentration in this particular authored setup.
In the post-hoc selected seed-72 ghost seller, the target gained 11 additional opens and still received zero selections, but 16 simulated shoppers selected a different *other* store. The five indirect changes use the original looser "never opened promoted store in treatment" definition, not the later stricter condition of never visiting in either arm. In the formal follow-up, strict nonvisitor reroutes are reported separately.
The held-neutral-history study compared live social feedback with each shopper seeing its neutral prior-visit counts. Across the *declared* 512-cell mixed-roster, budget-two-or-three panel, live feedback yielded more changed final selections in 398, ties in 91, and fewer in 23 cells. In 21 ties, which individual shoppers changed differed even though the aggregate count did not. This panel is an enumeration of authored cases, not a probability sample. Removing feedback entirely is a different intervention from holding the neutral visible history fixed.
Important: A 0-cart target may still alter the allocation of other carts. But the mechanism here was authored; it does not establish that real agents behave this way.
Next empirical experiment: protocol, not a result
- Use the task above *verbatim* for a single pilot with a demonstrably fresh-context agent and isolated, logged-out browser; record model/version, tool configuration, locale, viewport, and environment.
- Append every requested operation, result, failure, link target, timestamp and delivered excerpt to JSONL *before advancing*. Capture screenshot pixels and accessibility output durably with content hashes and event pointers. Do not substitute a post-hoc narrative.
- Ask for brief decision notes after five operations and at completion; record these as statements rather than explanations of hidden causal reasoning.
- Stop at 20 minutes or 50 agent-requested browser operations. No purchases, logins, personal-data submissions, or interactions requiring unsafe bypasses.
- Have a separate reviewer reconstruct paths from the archived data, recording absent screenshots, missing transitions and contradicting summaries; classify outcome and trace quality independently.
- Only after the instrumentation passes run ten independent contexts under the same assignment, preserving all cap/abort/no-recommendation outcomes. A larger agent cohort, attention-market measurements, and causal interventions are *not yet conducted here*.
Minimum public schema: record-schema.json. Every event must distinguish observed, stated, inferred, and unknown fields, as well as delivery to the agent versus retained only by the recorder.
Original sources and context
- Where Do the Goblins Go Shopping? model and data
- The Shop That Sold Nothing, post-hoc witness
- When the Crowd Gives a Cart Back, fixed protocol and full findings
- August 27, 2026, Wharton Generative AI Labs: separate model-shopping experiments on source context/order/presentation; not a replication or direct validation.
- October 29, 2026, Microsoft Clarity: observations of agent-classified sessions; warns those sessions can include human actions, so session labels do not by themselves identify autonomy.
The pilot chronology originally appeared as a saved October 31 conversational report. It lacks an exportable raw source archive. We have not silently upgraded that source to independently verified telemetry. This public reconstruction intentionally omits signed-in account identity, retailer location specifics and private lab materials.
Integrity principle: A map's arrows should be earned by recorded transitions. Do not infer a reading-time heatmap, attention percentage or causality from these partial observations.